A generated title card reading 'Qwen 4 vs GPT-6.1 Sol: the model that costs a third as much per point', with the subtitle '2.3 cents per index point against 7.6, and a 272,000-token cliff that reprices the whole request'. Three numbered rows read 'index score 42.9 against 54.6, Artificial Analysis v5.0.1', 'cost per completed index task $3.27 against $1.27' and 'input doubles from $2.00 to $4.00 above 272,000 tokens'. A footer reads 'Scores and per-task cost per Artificial Analysis Intelligence Index v5.0.1; prices per each vendor's published rate card.' Flat vector styling on white with blue-and-cyan accents, the OrcaRouter mark composited bottom right.
Guides & Insights

Qwen 4 vs GPT-6.1 Sol: The Model That Costs a Third as Much Per Point

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Every other Qwen4-Max comparison in this series has the same shape: the opponent scores higher and costs more, and the argument is about whether the difference is worth the premium. This one does not. GPT-6.1 Sol scores 54.6 on the Artificial Analysis Intelligence Index against Qwen4-Max's 42.9 — twelve points ahead — and completes an index task for $1.27 against Qwen4-Max's $3.27. Sol is more capable and 2.6 times cheaper per finished answer at the same time. On the efficiency measure that decided the other four matchups, Sol wins by a factor of 3.3: 2.3 cents per index point against 7.6.

There is one condition attached to that, and it is the 272,000-token line. It is a real cliff, it applies to every token in the request, and it is the only route back to Qwen4-Max in this matchup that does not involve paying more for less.

The two cards

GPT-6.1 Sol shipped on 29 September 2026 at $2.00 in and $10.00 out per million tokens. Qwen4-Max went metered on 16 November 2026 at $1.50 and $4.50 in Singapore. On the headline rate Qwen4-Max looks 2.2× cheaper on output; per completed task it is 2.6× dearer. That inversion is the article.

• Input, per million — Sol $2.00 up to 272K tokens, $4.00 above it. Qwen4-Max $1.50 in Singapore, $1.413 in the other five regions, flat across the full window.

• Output, per million — Sol $10.00 short-context, $15.00 long-context. Qwen4-Max $4.50 and $4.24, flat.

• Cached input — Sol $0.10 per million short-context, $0.20 long; Qwen4-Max $0.075 reads, with $1.875 and $3.00 write rates for five-minute and one-hour entries.

• The cliff — O​penAI reprices every token in a request once input crosses 272,000, not just the excess. Below it you are on the cheap card; above it you are not, and there is no partial application.

• Context — Sol publishes a 1,000,000-token window on the tracker; Qwen4-Max the same, with thinking input capped at 983,616.

• Cost per index task — $1.27 versus $3.27.

• Index score — 54.6 versus 42.9, both measured at their top published effort setting.

A generated two-column scoreboard titled 'Qwen4-Max vs GPT-6.1 Sol - the scoreboard'. The Qwen4-Max column reads index score 42.9, cost per index task $3.27, cents per index point 7.6, input per million $1.50, output per million $4.50, hallucination rate 28.4%. The GPT-6.1 Sol column reads index score 54.6, cost per index task $1.27, cents per index point 2.3, input per million $2.00, output per million $10.00, hallucination rate 51.9%. A footer reads 'Scores and per-task cost per Artificial Analysis Intelligence Index v5.0.1; prices per each vendor's published rate card.' The OrcaRouter mark is composited bottom right.

Why the cheap-looking card is the expensive one

The per-task numbers only make sense once you look at what each model spent to get there. Sol emitted 63.7 million output tokens across the index suite; Qwen4-Max emitted 127.1 million. Qwen4-Max wrote almost exactly twice as much output to score twelve points lower, and output is the expensive column on both cards.

Per task the split is 67,824 output tokens for Sol against 104,940 for Qwen4-Max — of which 49,213 and 68,914 respectively are reasoning. So Qwen4-Max is not chattier in its answer; it is chattier in its thinking. It reasons 40% longer to reach a worse conclusion, and then bills you at $4.50 a million for the privilege.

Cache behaviour makes it worse rather than better. Qwen4-Max's $3.27 is 86% input, and $1.67 of that input is cache writes at $1.875 per million — a rate that exists because A​libaba charges for populating the cache, not just for reading it. Sol's $1.27 is 47% input, with $0.33 of cache writes and $0.25 of cache reads. On a workload with heavy prefix reuse, Qwen4-Max's cache-write line is the one that grows fastest, and it is the line most people do not model.

Where Qwen4-Max is not beaten

Twelve points is not a rout on every suite, and three rows are worth reading before writing the model off.

• Terminal-Bench 4.0 — Sol 56.1%, Qwen4-Max 48.5%. Seven and a half points. Real, but narrower than the headline suggests.

• AutomationBench — Qwen4-Max 72.6% against Sol's 64.9%. Qwen4-Max wins this one by nearly eight points, and it is the largest single inversion in the comparison.

• GDPval — Qwen4-Max 1,808.69 against Sol's 1,607.47. A 12.5% lead on economically grounded professional tasks, against a model twelve index points ahead. This is the second inversion and the more surprising of the two, because GDPval is the closest thing on the board to "would a professional accept this work".

• FrontierCode 2.0 — 26.5% versus 40.6%. Thirteen points the other way; the widest gap on the board and the most commercially loaded one.

• CritPt Agentic Hard — 43.0% against 69.5%. Twenty-six points, the worst single row for Qwen4-Max.

• Omniscience — 27.7 against 42.0, with hallucination rates of 28.4% and 51.9%. Sol is more accurate and much more likely to fabricate when wrong.

• Capability sub-scores — Sol leads all six, narrowly in strategy and operations (57.6 against 56.9) and clearly in engineering (59.4 against 50.1).

The pattern is consistent with the other matchups in this series: Qwen4-Max holds up where the task is professional judgement applied to documents, and loses where the task is sustained technical reasoning or agentic execution. Two wins out of a dozen suites is not a reason to choose it, but it is a reason the choice is task-dependent rather than automatic.

A headless-browser screenshot of the Artificial Analysis page for GPT-6.1 Sol, captured 17 November 2026 in English, showing OpenAI's mark, the label Proprietary model, Released September 2026, the heading GPT-6.1 Sol: Intelligence, Performance & Price Analysis, a rank strip reading Intelligence #12 of 81, Performance #62 of 79, Cost #21 of 80 and Verbosity #35 of 80, a statistics row giving a time to first answer token of 161.08 seconds, 49 output tokens per second, $2.00 input and $10.00 output per million tokens, a 95% cache discount, 1.92 billion input tokens and an output-to-input ratio of 3.322%, and the summary paragraph describing the model as among the leading models in intelligence and reasonably priced for its band, notably slow per task but fairly concise, with text and image input, text output and a 1M-token context window.

Running the 272,000-token arithmetic

Take a 400,000-token codebase you want analysed in one shot, with 2,000 tokens of output. Both models fit it inside their windows. Only one of them reprices.

• GPT-6.1 Sol, short-context rate — cannot be used. At 400K input the long-context tier applies: 400,000 × $4.00 per million = $1.60 input, plus 2,000 × $15.00 = $0.03 output. Total $1.63.

• Qwen4-Max — 400,000 × $1.50 per million = $0.60 input, plus 2,000 × $4.50 = $0.009 output. Total $0.61.

• The same request at 250,000 tokens, under the cliff — Sol: 250,000 × $2.00 = $0.50 plus $0.03 = $0.53. Qwen4-Max: $0.375 plus $0.009 = $0.38.

So the reversal is real and it is narrow. Below 272,000 input tokens Sol is cheaper per request and better; above it, Qwen4-Max is 2.7× cheaper and twelve points worse. A 200K-token agent loop with tool output accumulating across turns crosses the threshold partway through a session, which is exactly the shape of workload where the cliff bites hardest — you do not choose the tier, the loop's context growth chooses it for you.

These are list-price arithmetic on published rates, not measured invoices. Cache hits would cut both figures substantially and would narrow the absolute gap while leaving the ratio roughly where it is.

The routing answer, which is unusually clean here

GPT-6.1 Sol is on our catalogue as openai/gpt-6.1-sol at O​penAI's list price, including both context tiers, with 0% markup on top. Qwen4-Max is not on our catalogue — a qwen4 identifier returns not-found, its route is A​libaba's Model Studio API and several third-party platforms, and the Qwen model we do serve is qwen/qwen3.8-max at $2.00 and $6.00.

What makes this pairing worth routing rather than choosing is that the decision variable is a token count, and a routing layer can see it. A 272,000-token rule on the input length — Sol below, something cheaper above — is a two-line policy, and it is the kind of thing that is easy to implement once and impossible to remember to do by hand in every service. Automatic failover covers the other edge: Sol's long-context tier is a different price band, not a different model, so a spike in prompt size produces a bill change rather than an error, and nothing warns you until the invoice.

A headless-browser screenshot of the OrcaRouter model page for GPT-6.1 Sol, captured 17 November 2026 in English, showing the identifier openai/gpt-6.1-sol, the vision, tools, JSON and reasoning badges, the byline OpenAI dated 2026-09-29, a description placing it one tier below GPT-6 Astra and aimed at agentic coding and computer use, a 1M-token context, 128K max output, text, image and file input, and best-for tags for reasoning, coding and agentic work.

What would move this

Two things. A Qwen4-Max price cut would need to reach $0.60 in and $1.80 out before it undercut Sol on the short-context tier per task, which is a 60% reduction and not on any published roadmap. And a Sol long-context rate that applied to the excess rather than the whole request would remove the only structural advantage Qwen4-Max holds here; O​penAI's documentation is explicit that it does not, so that is a hypothetical rather than a forecast.

Until either happens, the rule is short and it does not flatter the subject of this series. For nearly everything, GPT-6.1 Sol is the better model and the cheaper one, and the twelve-point gap is being sold at a discount. For the specific case of a document-heavy professional-judgement task where GDPval's 12.5% lead means something, Qwen4-Max is worth a look at 2.6× the price per task. And for anything that drags a large context past 272,000 tokens, Qwen4-Max stops being the compromise option and becomes the correct one — not because it improved, but because the alternative doubled.